Why Medicine Needs Neuro-symbolic AI

In March, a 63-year-old woman goes to her doctor for a routine checkup. A blood test shows that her blood sugar has been running high, and she is told she has type 2 diabetes. She starts a common medicine for it and is asked to return in three months.

In June, she has her blood drawn at a lab near her office, which is easier than driving to the clinic. Her sugar is higher than it was in March. At the follow-up visit the nurse weighs her and finds she has lost about nine pounds, and the note records that she is responding well to lifestyle changes. She and her doctor are both pleased.

In August, she visits an urgent care clinic with an ache in the middle of her back that has been there for a few weeks. She has not fallen or lifted anything heavy. She is given an anti-inflammatory and referred to physical therapy.

Each of the three doctors did the job in front of them, and each record, read on its own, looks reasonable. Read together, they point to something none of them shows alone, and a doctor who saw all three would want a scan of her pancreas.

Why the three visits belong together

The pancreas is a small organ that sits deep behind the stomach, against the back wall of the abdomen. One of its jobs is to make insulin, the hormone that keeps blood sugar in check.

This is what links the three visits. A tumor in the pancreas can disturb how insulin is made, so blood sugar rises, sometimes before any other sign appears. Cancer can also cause weight loss that the person did not try for. And because the pancreas lies close to the spine and its nerves, a tumor there can cause a dull pain in the middle of the back.

Most people with new diabetes do not have cancer. When diabetes appears for the first time after the age of 50, however, about one in a hundred are diagnosed with pancreatic cancer within three years, roughly eight times the rate expected at that age. Researchers at the Mayo Clinic built a score to find these patients, based on the age at which diabetes began, the change in weight and the change in blood sugar. Weight that falls while sugar rises counts the most.

The timing matters because pancreatic cancer is usually found late. About half of cases are diagnosed only after the cancer has spread, when treatment can do much less. Finding it months earlier can change what is possible.

In this woman's case, the information needed was already recorded. It was spread across three places, and no one was looking at all of it.

The same gap works in the other direction. I have seen patients go through cardiac evaluations when the cause was anxiety and the irregular heartbeat was its symptom, and I have treated patients like that. Their hearts were checked with care, but the answer lay elsewhere in their story. When the pieces of a person's health sit apart, medicine misses what is serious and chases what is not.

Why computers have not solved this

Computers read far faster than people, so it is fair to ask why they do not already catch cases like this. The reason is that most AI in medicine is one of two kinds, and each sees only part of the picture.

The first kind learns from examples. Given millions of past patients, it learns which patterns tended to come before an illness. The chatbots many people now use are built this way, trained on very large amounts of text. This kind of AI is good at noticing, but it cannot show why it reached a conclusion, and it sounds as confident when it is wrong as when it is right. A doctor cannot safely act on a conclusion she has no way to check.

The second kind follows rules that people have written down, in the form "if this, then that." A rule is clear about what it checks and why. It catches only what someone thought of in advance, however, and only when the information arrives in exactly the form the rule expects. It does not learn.

Each approach is partial, and neither is a bad idea.

A deeper problem lies in the records these systems learn from. Much of a medical record consists of notes, and each note is one person's account of a visit, written quickly and shaped by how that person saw the patient. Much of the rest was designed for billing, with diagnoses stored as payment codes. A system that learns only from this material learns its blind spots as well. In 2019, researchers found that an algorithm used by American health systems to choose patients for extra care was judging Black patients to be healthier than they were. It had learned from past spending, and less had been spent on their care.

Learning and logic together

Good doctors use both kinds of thinking. Years of practice teach them to notice when something is off, and they then check that impression against the evidence before they act. Neuro-symbolic AI is built in the same way. The "neuro" part learns and notices, as experience does. The "symbolic" part reasons with rules, as a doctor does when checking against the evidence.

At Chroneme, the rules come from published research and medical guidelines, and they work like a lens. For each clinical question, they set out which information matters and how far to trust its source, so that a weight measured on a scale counts for more than a word in a note. Within that frame, the learning part can still notice patterns that no one thought to write down. The logic part then tests each one against the evidence, and a pattern that does not hold up never reaches the doctor as a recommendation.

Consider the same woman with Chroneme reading her care. Her lab results, weights and notes are brought into one record, from every place she was seen, in the order the events happened. In June, when her sugar rises while her weight falls, the pattern matches what the published score looks for, and her doctor sees the finding that week, two months before the back pain began. The finding is written in plain language and shows every result behind it, where and when each was recorded, and the study the rule is drawn from. A scan order is already prepared, along with a message to the patient that explains the reason without alarming her. Her doctor reviews it, agrees or disagrees, and signs.

Because the rules are written down, the same facts lead to the same conclusion for every patient, regardless of who is on call. When the evidence changes, the rule is updated once and applies to everyone. When something goes wrong, the point of failure can be found.

What we are building

A better method is not enough on its own. It needs a system to run on, one that keeps each person's record whole and works as well for the millionth patient as for the first. That system is what we are building.

It also makes new kinds of insight possible. A scan shows what the body looks like inside, a history shows how it got there, and genetics shows what it was more likely to do. These have usually been read separately, by different specialists. Read together and in order, they reveal combinations that no single record contains.

Medicine has known for decades what good, evidence-based care looks like. What it has lacked is the time to apply all of that evidence to every patient, every day, since no doctor can review every record against every guideline before each visit. Chroneme's agents, programs that carry out the steps of a task, do that preparation. They draft the orders, messages and notes, each with its sources, for a clinician to review and sign. The judgment stays with people, and the time saved goes back to care.

We are careful about what we claim. That every conclusion can be traced and checked is true by design today. Whether the system is right often enough to help patients is a question for evidence. We are testing it under a protocol written before any results were seen, and we will publish what we find, including where it falls short.